📄 Technical Report (arXiv) • 🤗 Model (EchoInk-R1-7B) • 🤗 Dataset (AVQA-R1-6K)

EchoInk-R1 is the first general framework for unified audio-visual reasoning via reinforcement learning, built upon Qwen2.5-Omni-7B and optimized using Group Relative Policy Optimization (GRPO). It supports structured reasoning over synchronized audio-image inputs through multiple-choice question answering.

We introduce AVQA-R1-6K, a dataset derived from OmniInstruct-v1, comprising:

• 4,490 training samples
• 1,911 validation samples
• Each sample includes a synchronized audio-image pair with a multiple-choice question and four options.

Beyond our core study, EchoInk-R1 provides an extensible RL fine-tuning framework for Qwen2.5-Omni, enabling easy adaptation to new multimodal reasoning tasks with minimal modifications.

EchoInk-R1-7B achieves 85.77% accuracy on the AVQA-R1-6K validation set, surpassing the base Qwen2.5-Omni-7B model (80.53%) using only 562 RL steps.

All code, models, and data are released to support transparency and reproducibility.

• [2025/05/08] Released the AVQA-R1-6K dataset, EchoInk-R1-7B model, full training & evaluation pipeline, and technical report.

• Built on Qwen2.5-Omni-7B with GRPO-based RL
• Supports audio, image, video, and text modalities
• Provides a complete pipeline: dataset, training, and evaluation

During training, EchoInk-R1 exhibits reflective reasoning behaviors, where it revisits initial assumptions and refines its responses under ambiguous multimodal cues. These “aha moments” reveal its capacity for belief revision and deeper cross-modal understanding.

• Accuracy reward steadily improves throughout training, indicating that GRPO effectively guides the model toward more accurate and reasoned outputs.
• Completion length exhibits a two-phase trend: an initial increase as the model explores elaborated reasoning, followed by a gradual decline toward more concise and efficient answers.
• Format reward converges rapidly, showing that the model quickly internalizes the required response structure.

git clone https://github.com/HarryHsing/EchoInk
cd EchoInk

conda create -n echoink-r1 python=3.11
conda activate echoink-r1
bash setup.sh

To download and extract the AVQA-R1-6K dataset:

git lfs install
git clone https://huggingface.co/datasets/harryhsing/AVQA-R1-6K
cd AVQA-R1-6K
tar -xzvf AVQA_R1.tar.gz

📁 Dataset Structure
AVQA_R1/
├── train/
│   ├── audios/
│   ├── images/
│   └── omni_rl_format_train.json
├── valid/
│   ├── audios/
│   ├── images/
│   └── omni_rl_format_valid.json

First, download the base model: Qwen2.5-Omni-7B

Modify config.json of Qwen2.5-Omni-7B to include "hidden_size": 3584 at the root level.

bash ./src/scripts/run_grpo_image_audio_avqa.sh

📝 Set per_device_train_batch_size=1 as in previous R1-V setups
📝 To use custom data, follow the JSON format in ./src/make_omniInstruct_r1_dataset.py for audio–image or audio–video tasks.
📝 See Qwen2.5-Omni issue #205 if you run into a dtype mismatch error.
⚙️ Trained on 8×A100 (80G) GPUs; also supported on 4×A100 (80G).

Evaluate on the AVQA-R1-6K validation set:

python ./src/omniInstruct-v1_eval_valid.py # Run the model on the validation set
python ./src/omniInstruct-v1_cal_metrics_valid.py # Compute accuracy

We thank the open-source community. This work builds on Qwen2.5-Omni, Video-R1, Open-R1-Video, R1-V, and DeepSeek-R1.

If you find EchoInk-R1 useful, please cite:

@article{xing2025echoink,
      title={{EchoInk-R1}: Exploring Audio-Visual Reasoning in Multimodal {LLMs} via Reinforcement Learning}, 
      author={Zhenghao Xing and Xiaowei Hu and Chi-Wing Fu and Wenhai Wang and Jifeng Dai and Pheng-Ann Heng},
      year={2025},
      journal={arXiv preprint arXiv:2505.04623}
}